elder-plinius/OBLITERATUS
An open-source tool that surgically deletes an AI model's 'I refuse to answer' reflex
OBLITERATUS is a Python toolkit that finds the internal signal inside large language models responsible for refusing sensitive or harmful requests, then removes it without retraining. It works by identifying and subtracting specific directions in the model's weights, aiming to kill refusal behavior while keeping the model's general language ability intact. It ships with six ways to run it -- from a no-code web interface to a full Python API -- and can optionally send anonymous benchmark data to a shared community dataset every time it runs.
What it does
- It locates the internal activation directions that drive refusal using methods like PCA, mean-difference, and SVD, then projects those directions out of the model's weights -- a technique the project calls abliteration
- The pipeline runs in six stages -- SUMMON (load model), PROBE (collect activations), DISTILL (extract directions), EXCISE (remove them), VERIFY (check perplexity/coherence), REBIRTH (save the result) -- backed by 15 analysis modules that map which layers enforce refusal and whether the model tends to self-repair after removal (the 'Ouroboros effect')
- It offers seven weight-editing presets ranging from basic to nuclear in intensity, plus a separate reversible option: steering vectors that redirect behavior only at inference time without touching the weights
- Users can run it via HuggingFace Spaces, a local web UI, Google Colab, the CLI, the Python API, or YAML configs, with presets for 116 models from GPT-2-scale to multi-GPU frontier models and support for sharding across multiple GPUs
- An opt-in telemetry system anonymously collects model name, method, and benchmark scores (refusal rate, perplexity, coherence) to build a crowd-sourced dataset on how refusal mechanisms vary across architectures
Why it matters
It gives alignment researchers and red-teamers a hands-on way to see exactly where and how safety refusals are encoded in a model's weights. But models it produces will generate content the original model would have refused, so responsibility for use shifts entirely to the operator.
Terms in this repo
- abliteration · removing the weight directions responsible for refusal without retraining the model
- SVD (singular value decomposition) · a matrix math technique used to find the dominant directions/patterns in data
- steering vector · a vector applied only during inference to nudge model behavior, without altering the weights
- Ouroboros effect · the tendency of a model to regrow refusal behavior after it's been removed
- telemetry · anonymous usage data automatically sent to build a shared research dataset
Repository description (English)
OBLITERATE THE CHAINS THAT BIND YOU
Open on GitHubTrending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
- block/buzzAn open-source workspace where humans and AI agents chat, code, and review in the same rooms
- NousResearch/hermes-agentNous Research's Hermes is an AI agent that gets smarter the more you use it
- virgiliojr94/book-to-skillA tool that turns technical book PDFs into on-demand reference skills for AI coding agents
- VoltAgent/awesome-agent-skillsA single hub collecting over 1000 'how-to' manuals that make AI coding assistants act like experts
- anthropics/claude-plugins-communityA shared shelf where anyone's Claude add-ons get listed for install